Research Of Live Virtual Machine Migration Based On Dirty Page Probability Forecasting

Memory migration transfers the largest amount of data in live virtualmachine migration. At present, iterative Pre-copy mechanism is often used tosolve the challenges of memory migration. Since the VM is still running duringmigration, memory pages might be modified and these dirty pages should bere-transmitted for consistency. The frequent re-transmission of high dirty pagesleads to a dramatic decline of migration efficiency. As a result, it is an urgentproblem to reduce data transferred and migration time.This thesis focuses on memory migration of live VM migration. Firstly, itpresents a new Pre-copy approach based on dirty page probability forecasting.According to the Temporal Locality, high dirty pages are being writtenfrequently in a short period, meeting the stationary condition of autoregressiveprocess. The new approach utilizes the AR(n) model, which is based on the timesequence, to forecast dirty probability of memory pages, and distinguishes highdirty pages from dirty pages, and then prefers to transfer pages with a lowerdirty probability. Such approach can avoid the re-transmission of high dirtypages, and contribute to a better performance of live migration. Secondly, we implement live VM migration based on dirty page probabilityforecasting in Xen and test the performance of the new method under differentCPU loads, memory loads and network loads. And we also analyze theexperiment results in three different aspects, including total migration time,downtime and data transmitted. The results demonstrate that the new approachreduces data transferred, decreases total migration time and thus effectivelyimproves the performance of live VM migration.Finally, this thesis applies live VM migration technology ontoself-developed "Information Security Experiment Platform", designs andrealizes the ISEP live migration system which can schedule the resource of theplatform’s virtual machine cluster. The ISEP optimizes the resource’quantitativemethod of VMs and their hosts, brings in K/N model to avoid frequent migrationaroused by transient peak load, and combines the lowest-cost principle and thebest-fit principle to choose the source VM and the destination host. The ISEPwill achieve both the goals of load balancing and energy saving.